| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Message-ID: <387528a7-42b3-4576-a03a-448c1a90f9df@paulmck-laptop>
Date: Wed, 4 Oct 2023 09:07:37 -0700
From: "Paul E. McKenney" <paulmck@...nel.org>
To: Uladzislau Rezki <urezki@...il.com>
Cc: RCU <rcu@...r.kernel.org>, quic_neeraju@...cinc.com,
Boqun Feng <boqun.feng@...il.com>,
Joel Fernandes <joel@...lfernandes.org>,
LKML <linux-kernel@...r.kernel.org>,
Oleksiy Avramchenko <oleksiy.avramchenko@...y.com>,
Frederic Weisbecker <frederic@...nel.org>
Subject: Re: [PATCH v2 1/1] rcu: Reduce synchronize_rcu() waiting time
On Wed, Oct 04, 2023 at 05:26:52PM +0200, Uladzislau Rezki wrote:
> On Wed, Oct 04, 2023 at 08:24:35AM -0700, Paul E. McKenney wrote:
> > On Tue, Oct 03, 2023 at 08:04:03PM +0200, Uladzislau Rezki (Sony) wrote:
> > > A call to a synchronize_rcu() can be optimized from time point of
> > > view. Different workloads can be affected by this especially the
> > > ones which use this API in its time critical sections.
> > >
> > > For example if CONFIG_RCU_NOCB_CPU is set, the wakeme_after_rcu()
> > > callback can be delayed and such delay depends on where in a nocb
> > > list it is located.
> > >
> > > 1. On our Android devices i can easily trigger the scenario when
> > > it is a last in the list out of ~3600 callbacks:
> > >
> > > <snip>
> > > <...>-29 [001] d..1. 21950.145313: rcu_batch_start: rcu_preempt CBs=3613 bl=28
> > > ...
> > > <...>-29 [001] ..... 21950.152578: rcu_invoke_callback: rcu_preempt rhp=00000000b2d6dee8 func=__free_vm_area_struct.cfi_jt
> > > <...>-29 [001] ..... 21950.152579: rcu_invoke_callback: rcu_preempt rhp=00000000a446f607 func=__free_vm_area_struct.cfi_jt
> > > <...>-29 [001] ..... 21950.152580: rcu_invoke_callback: rcu_preempt rhp=00000000a5cab03b func=__free_vm_area_struct.cfi_jt
> > > <...>-29 [001] ..... 21950.152581: rcu_invoke_callback: rcu_preempt rhp=0000000013b7e5ee func=__free_vm_area_struct.cfi_jt
> > > <...>-29 [001] ..... 21950.152582: rcu_invoke_callback: rcu_preempt rhp=000000000a8ca6f9 func=__free_vm_area_struct.cfi_jt
> > > <...>-29 [001] ..... 21950.152583: rcu_invoke_callback: rcu_preempt rhp=000000008f162ca8 func=wakeme_after_rcu.cfi_jt
> > > <...>-29 [001] d..1. 21950.152625: rcu_batch_end: rcu_preempt CBs-invoked=3612 idle=....
> > > <snip>
> > >
> > > 2. On our Android devices we use cpuset/cgroup to classify tasks
> > > and assign them into different cgroups. For example "backgrond"
> > > group which binds tasks only to little CPUs or "foreground" that
> > > binds to all CPUs, i.e. tasks can be migrated between groups.
> > >
> > > See below an example of how "surfaceflinger" task is migrated.
> > > Initially it is located in the "system-background" cgroup which
> > > allows to run only on little cores. In order to speedup it up
> > > it can be temporary moved into "foreground" cgroup which allows
> > > to use big CPUs:
> > >
> > > cgroup_attach_task():
> > > -> cgroup_migrate_execute()
> > > -> cpuset_can_attach()
> > > -> percpu_down_write()
> > > -> rcu_sync_enter()
> > > -> synchronize_rcu()
> > > -> now move tasks to the new cgroup.
> > > -> cgroup_migrate_finish()
> > >
> > > <snip>
> > > rcuop/1-29 [000] ..... 7030.528570: rcu_invoke_callback: rcu_preempt rhp=00000000461605e0 func=wakeme_after_rcu.cfi_jt
> > > PERFD-SERVER-1855 [000] d..1. 7030.530293: cgroup_attach_task: dst_root=3 dst_id=22 dst_level=1 dst_path=/foreground pid=1900 comm=surfaceflinger
> > > PERFD-SERVER-1855 [000] d..1. 7030.530383: cgroup_attach_task: dst_root=3 dst_id=22 dst_level=1 dst_path=/foreground pid=1900 comm=surfaceflinger
> > > TimerDispatch-2768 [002] d..5. 7030.537542: sched_migrate_task: comm=surfaceflinger pid=1900 prio=98 orig_cpu=0 dest_cpu=4
> > > <snip>
> > >
> > > from this example it is clear that "a moving time" also depends
> > > on how fast synchronize_rcu() completes.
> > >
> > > 3. This patch improves the synchronize_rcu() approximately by 30%-50%
> > > on synthetic tests. Apart of that i have tested app launch of camera
> > > app where i also see better perf. figures:
> > >
> > > 542 vs 489 diff: 9%
> > > 540 vs 466 diff: 13%
> > > 518 vs 468 diff: 9%
> > > 531 vs 457 diff: 13%
> > > 548 vs 475 diff: 13%
> > > 509 vs 484 diff: 4%
> > >
> > > Signed-off-by: Uladzislau Rezki (Sony) <urezki@...il.com>
> > > ---
> > > kernel/rcu/tree.c | 151 +++++++++++++++++++++++++++++++++++++++++-
> > > kernel/rcu/tree_exp.h | 2 +-
> > > 2 files changed, 151 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> > > index 78554e7181dd..a347c1f98f11 100644
> > > --- a/kernel/rcu/tree.c
> > > +++ b/kernel/rcu/tree.c
> > > @@ -1384,6 +1384,122 @@ static void rcu_poll_gp_seq_end_unlocked(unsigned long *snap)
> > > raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
> > > }
> > >
> > > +/*
> > > + * There are three lists for handling synchronize_rcu() users.
> > > + * A first list corresponds to new coming users, second for users
> > > + * which wait for a grace period and third is for which a grace
> > > + * period is passed.
> > > + */
> > > +static struct sr_normal_state {
> > > + struct llist_head curr; /* request a GP users. */
> > > + struct llist_head wait; /* wait for GP users. */
> > > + struct llist_head done; /* ready for GP users. */
> > > + struct llist_node *curr_tail;
> > > + struct llist_node *wait_tail;
> > > + atomic_t active;
> > > +} sr;
> > > +
> > > +/* Enable it by default. */
> > > +static int rcu_normal_wake_from_gp = 1;
> > > +module_param(rcu_normal_wake_from_gp, int, 0644);
> >
> > Nice!
> >
> > But could you please make this default to zero in order to avoid
> > surprising people for whom the old way works better?
> >
> Yep, that i can do. If people prefer a slower version of it :)
Keeping in mind that "slower" often means more updates to spread the
grace-period overhead across, thus lower per-update RCU overhead. ;-)
Thanx, Paul
Powered by blists - more mailing lists